Bria FIBO 2.0 Edit Turbo (4-bit MLX) on Apple Silicon

Capability 1 · Natural Language Instruction

Single Image Edit: Text-Guided Attribute Transformation

Transform specific garments, attributes, lighting, or materials from a single reference image without altering identity, pose, or background depth. The integrated 4-bit Qwen3-VL vision encoder visually grounds the source while the 44-layer DiT executes the edit in only 4 Turbo steps.

Prompt: "the jacket color is now red"

Input Reference Image - Man in Yellow Jacket

Input Reference Source Image

Bria AI Reference bf16 Baseline

DiffusionBear Studio MLX 4-bit (4 steps)

Capability 2 · Multimodal In-Context Binding

Multiple References: Multi-Subject & Garment Transfer

Combine up to 4 separate images into a single coherent generation. In this benchmark, FIBO2 extracts the identity and pose of the woman from Reference 1, and maps her into the dress cut, color, and fabric texture from Reference 2 while synthesizing a pristine photographic scene.

Prompt: "Generate a new image of the woman from Image 1, keeping her exact face, hairstyle, pose, body shape, and background from Image 1, but make her wear the exact dress from Image 2 (matching the same color, cut, and fabric)."

Reference 1 (Identity) Input A

Reference 2 (Garment) Input B

Bria AI Reference bf16 Baseline

DiffusionBear Studio MLX 4-bit (4 steps)

Capability 3 · Sub-Pixel Canvas Precision

Masked Edit (Inpainting): Zero-Variance Area Replacement

Paint directly on the reference backdrop in DiffusionBear Studio. Unmasked regions are strictly preserved with deterministic 1.5px Gaussian edge feathering and alpha compositing, eliminating ambient variance outside the masked area while replacing the target subject with surgical fidelity.

Prompt: "replace the suitcase with a vintage brown leather one"

Original Scene Source Reference

Studio Painted Mask Binary Mask

Bria AI Reference bf16 Baseline

DiffusionBear Studio MLX 4-bit (4 steps)

Engineering Telemetry

Quantization & Memory Footprint on Apple Silicon

Stock bf16 FIBO 2.0 weights require over 24 GB of storage and peak above 17.4 GB in VRAM, causing immediate out-of-memory aborts or severe OS memory thrashing on 16 GB Apple Silicon Macs. DiffusionBear's affine 4-bit MLX quantization compresses the pipeline to 7.33 GB on disk and maintains a runtime resident working set strictly under 6.8 GB.

Component	Original Precision (bf16)	4-bit MLX (`group_size=64`)	Compression Ratio
DiT Transformer (44 layers, dim 3584)	15.32 GB	4.26 GB	-72.2%
Vision Text Encoder (Qwen3-VL-4B-Instruct)	8.27 GB	2.88 GB	-65.2%
Latent VAE (FLUX.2 AutoencoderKL fp16)	0.34 GB	0.19 GB	-44.1%
Total Model Size on Disk	24.3 GB	7.33 GB	-69.8%
Resident Working Set (VRAM)	17.4 GB (OOM / Thrash)	< 6.8 GB	Runs 100% in RAM
Inference Time (512×768, 4 steps, M1)	N/A (Disk swap death)	~85s (21s / step)	Zero Swap

⚖️

Bria AI FIBO 2.0 — Licensing, Legal Terms & Official HF Repo Notice

Official Licensing Notice from Bria AI (briaai/Fibo-2.0-edit-turbo):

“License: Creative Commons Attribution–Non-Commercial (CC BY-NC 4.0)
The model is released under a CC BY-NC 4.0 license for non-commercial use.
Commercial use is subject to a commercial agreement with BRIA. Available at platform.bria.ai/licensing/pricing?model=fibo-2.”

Official HF Gated Model Card Agreement:

“I understand that commercial use of this model requires a separate commercial agreement with BRIA AI, and that academic access is granted on request and is limited to non-commercial research and teaching.”

Non-Commercial License (CC BY-NC 4.0): The open-weights model and its quantized derivatives (including this 4-bit MLX port, Ouinche/FIBO2-Edit-Turbo-MLX-4bit) are released strictly for non-commercial evaluation, personal experimentation, educational, and academic research purposes. You may freely use, test, evaluate, and share the model within these non-commercial boundaries.
Commercial Use Restriction: Any commercial deployment — including integration into commercial SaaS applications, paid web services, fee-based creative tools, monetization, or client deliverables — is expressly prohibited under the open-weights release without a separate commercial agreement from Bria AI.
100% Licensed Training Data & Legal IP Safety: Unlike web-scraped foundation models, Bria AI's models are trained exclusively on fully licensed, curated datasets (~1 billion licensed images) with built-in attribution traceability and legal indemnity for enterprise clients. If your organization wishes to use FIBO 2.0 in a production or revenue-generating environment, contact Bria AI directly for commercial licensing:
👉 Bria AI Commercial Licensing & Enterprise Pricing | Visit bria.ai
Attribution & Provenance: In compliance with Section 3(a) of the Creative Commons Attribution-NonCommercial 4.0 International Public License, full credit is given to Bria AI (huggingface.co/briaai) as the creator of the original foundational weights and architecture.

Academic Citation (arXiv:2511.06876):

@article{gutflaish2025generating, title={Generating an Image From 1,000 Words: Enhancing Text-To-Image With Structured Captions}, author={Gutflaish, Eyal and Kachlon, Eliran and Zisman, Hezi and Hacham, Tal and Sarid, Nimrod and Visheratin, Alexander and Huberman, Saar and Davidi, Gal and Bukchin, Guy and Goldberg, Kfir and others}, journal={arXiv preprint arXiv:2511.06876}, year={2025} }

Bria FIBO 2.0 Edit Turboon Apple Silicon Metal

Single Image Edit: Text-Guided Attribute Transformation

Multiple References: Multi-Subject & Garment Transfer

Masked Edit (Inpainting): Zero-Variance Area Replacement

Quantization & Memory Footprint on Apple Silicon

Bria AI FIBO 2.0 — Licensing, Legal Terms & Official HF Repo Notice

Bria FIBO 2.0 Edit Turbo
on Apple Silicon Metal